Computer implemented method for digital image data, digital video data or digital audio data enhancement, and a computer implemented method for encoding or decoding this data in particular for transmission or storage, wherein an element representing a part of said digital data comprises an indication of a position of the element in an ordered input data of a plurality of data elements, wherein a plurality of elements is transformed to a representation depending on an invertible linear mapping, wherein the invertible linear mapping maps the input of the plurality of elements to the representation, wherein the invertible linear mapping comprises at least one autoregressive convolution.
Legal claims defining the scope of protection. Each claim is shown in both the original legal language and a plain English translation.
1. A computer implemented method for digital image enhancement, in which each element of a plurality of elements representing a pixel of a digital image includes an indication of a spatial dimension, the spatial dimension indicating a position of the pixel in the digital image, and the element includes an indication of a channel dimension, the channel dimension indicating a channel of the pixel in the digital image, the method comprising the following steps: transforming the plurality of elements representing pixels of the digital image to a representation depending on an invertible linear mapping, the invertible linear mapping mapping an input of the plurality of elements to the representation; modifying the representation to determine a modified representation depending on the representation; determining a plurality of elements representing pixels of an enhanced digital image depending on the modified representation; and transforming the modified representation depending on an inversion of the invertible linear mapping, wherein the invertible linear mapping includes at least one autoregressive convolution.
This invention relates to digital image enhancement, specifically improving image quality through a structured transformation process. The method addresses the challenge of enhancing digital images while preserving spatial and channel information. Each pixel in the input image is represented by an element containing spatial coordinates (indicating position) and channel data (indicating color or intensity values). The method transforms these pixel elements into a new representation using an invertible linear mapping, which includes at least one autoregressive convolution. This transformation allows for efficient manipulation of the image data while maintaining reversibility. The transformed representation is then modified to enhance the image, such as improving sharpness, contrast, or noise reduction. The modified representation is then converted back to the original spatial and channel dimensions using the inverse of the initial mapping, producing an enhanced digital image. The use of autoregressive convolution ensures that dependencies between pixels are preserved during transformation, improving the quality of the enhanced output. This approach enables high-quality image enhancement while maintaining computational efficiency and reversibility.
2. The computer implemented method as recited in claim 1 , wherein a plurality of digital images of a digital video are processed according to the method.
The invention relates to a computer-implemented method for processing digital images, particularly in the context of digital video. The method addresses the challenge of efficiently analyzing and extracting meaningful data from multiple digital images within a video sequence. The core technique involves applying a series of computational steps to each image in the video to enhance, filter, or otherwise process the visual data. These steps may include noise reduction, feature extraction, object detection, or other image processing techniques. The method ensures that the processing is consistent across all images in the video, maintaining temporal coherence and improving the overall quality or usability of the video data. By systematically applying these steps to each frame, the method enables applications such as video enhancement, automated analysis, or real-time processing in various domains like surveillance, medical imaging, or entertainment. The invention improves upon existing methods by providing a structured approach to handling multiple images in a video, ensuring efficiency and accuracy in the processing pipeline.
3. The computer implemented method according to claim 1 , wherein a convolutional neural network for the invertible linear mapping determines the representation from the input.
A computer-implemented method for processing data using a convolutional neural network (CNN) involves generating a representation of input data through an invertible linear mapping. The method addresses the challenge of efficiently transforming input data into a structured representation while preserving the ability to reverse the transformation. The CNN is specifically configured to perform this invertible linear mapping, ensuring that the representation can be accurately reconstructed back to the original input when needed. This approach is particularly useful in applications requiring both forward and inverse transformations, such as data compression, feature extraction, or signal processing. The method leverages the CNN's ability to learn spatial hierarchies in the data, making it suitable for tasks involving structured or grid-like inputs, such as images or time-series data. The invertible nature of the mapping ensures that no information is lost during the transformation, which is critical for applications where data integrity is paramount. The method may be applied in various domains, including machine learning, computer vision, and signal processing, where efficient and reversible data transformations are required.
4. The computer implemented method according to claim 1 , wherein the representation is determined depending on a first autoregressive convolution of the input and a first convolution filter, and depending a consecutive second autoregressive convolution of the first autoregressive convolution and a second convolution filter.
This invention relates to a computer-implemented method for generating representations of input data using a neural network architecture that employs autoregressive convolutions. The method addresses the challenge of efficiently processing sequential or structured data by leveraging recursive convolutional operations to capture dependencies across multiple layers. The method involves determining a representation of input data through a sequence of autoregressive convolutions. First, a first autoregressive convolution is applied to the input data using a first convolution filter, producing an intermediate output. This intermediate output is then processed by a second autoregressive convolution using a second convolution filter, generating the final representation. The autoregressive nature of the convolutions allows the model to iteratively refine the representation by incorporating feedback from previous convolutional layers, enhancing the ability to capture complex patterns in the data. The use of multiple autoregressive convolutions enables the model to learn hierarchical features, where each convolutional layer builds upon the output of the preceding layer. This approach improves the model's capacity to handle sequential dependencies and structured data, making it suitable for applications such as time-series analysis, natural language processing, or image recognition. The method optimizes computational efficiency by reusing intermediate results, reducing redundancy in feature extraction.
5. The computer implemented method according to claim 1 , wherein the autoregressive convolution imposes an order on the input such that values of the representation for a specific element depend only on elements of the input representing input that is in the imposed order before the specific element in the order.
This invention relates to a computer-implemented method for processing input data using autoregressive convolution, a technique in machine learning and signal processing. The method addresses the challenge of efficiently modeling sequential or ordered data, such as time-series signals or natural language, where dependencies between elements must be preserved in a structured manner. The method imposes an order on the input data, ensuring that the representation of any given element depends only on preceding elements in the defined order. This constraint enforces a causal or sequential dependency structure, which is critical for applications like predictive modeling, where future values should not influence past ones. The autoregressive convolution operation processes the input by applying a learned filter that respects this ordering, effectively capturing local dependencies while maintaining the imposed sequence. The method may be applied in various domains, including audio processing, natural language processing, and time-series forecasting, where maintaining temporal or sequential coherence is essential. By restricting dependencies to prior elements, the approach ensures that the model adheres to causal relationships, improving interpretability and performance in tasks requiring ordered data analysis. The technique can be integrated into neural network architectures, such as convolutional or recurrent networks, to enhance their ability to handle structured input data.
6. The computer implemented method according to claim 1 , wherein an input of an input dimension is mapped to the representation by a plurality of consecutive autoregressive convolutions, wherein a dimension of the consecutive convolutions is equal or less than the input dimension.
This invention relates to a computer-implemented method for processing input data using autoregressive convolutions in a neural network. The method addresses the challenge of efficiently transforming high-dimensional input data into a compact representation while preserving important features. The core technique involves applying a sequence of autoregressive convolutions, where each convolution operation processes the input data in a manner that depends on previous computations. The dimensionality of these convolutions is constrained to be equal to or less than the original input dimension, ensuring that the representation remains manageable while capturing relevant information. This approach is particularly useful in tasks requiring dimensionality reduction, feature extraction, or data compression, where maintaining computational efficiency is critical. The method leverages the autoregressive nature of the convolutions to iteratively refine the representation, allowing for progressive refinement of the output. The technique can be applied in various domains, including image processing, natural language processing, and time-series analysis, where high-dimensional data must be transformed into a lower-dimensional space without significant loss of information. The constrained dimensionality of the convolutions ensures that the method remains computationally feasible while still producing accurate and meaningful representations.
7. The computer implemented method according to claim 1 , further comprising the following step: determining a N-dimensional kernel for the mapping depending on concatenating a plurality of (N−1)-dimensional kernels with identical size one after another along the dimension N.
This invention relates to a computer-implemented method for constructing a high-dimensional kernel in machine learning or data processing applications. The method addresses the challenge of efficiently generating complex, multi-dimensional kernels for tasks such as feature mapping, dimensionality reduction, or kernel-based learning, where traditional approaches may suffer from computational inefficiency or scalability issues. The method involves determining an N-dimensional kernel by concatenating multiple (N−1)-dimensional kernels of identical size along the Nth dimension. Each (N−1)-dimensional kernel represents a lower-dimensional transformation or projection, and their sequential concatenation forms a higher-dimensional structure. This approach leverages the modularity of lower-dimensional kernels to construct a more complex, high-dimensional kernel without requiring explicit computation of the full N-dimensional space, thereby improving computational efficiency and scalability. The method is particularly useful in applications where high-dimensional mappings are needed but direct computation is impractical, such as in kernel methods for machine learning, signal processing, or computer vision. By decomposing the problem into manageable lower-dimensional components, the method enables the construction of sophisticated kernels while maintaining computational feasibility. The resulting N-dimensional kernel can then be used for tasks such as data transformation, classification, or regression in high-dimensional spaces.
8. The computer implemented method according to claim 7 , wherein determining the N-dimensional kernel includes associating the (N−1)-dimensional kernel to the N-dimensional kernel as a last dimension entry, wherein a size of the last dimension of the N-dimensional kernel defines a center value, wherein for any entries of the N-dimensional kernel in a last dimension of the N-dimensional kernel having an index smaller than the center value, arbitrary values are assigned, wherein for any entries in the last dimension having an index larger than the center value, zeros are assigned.
This invention relates to a computer-implemented method for constructing N-dimensional kernels in machine learning or signal processing applications. The method addresses the challenge of efficiently generating higher-dimensional kernels from lower-dimensional ones, particularly when the kernel must maintain specific symmetry or sparsity properties. The method involves determining an N-dimensional kernel by associating an (N−1)-dimensional kernel as the last dimension entry. The size of this last dimension defines a center value, which serves as a reference point for assigning values to the kernel entries. For entries in the last dimension with indices smaller than the center value, arbitrary values are assigned, while entries with indices larger than the center value are set to zero. This approach ensures that the kernel retains a structured sparsity pattern, which can be useful in applications requiring efficient computation or memory usage. The method can be applied iteratively, where an (N−1)-dimensional kernel is first constructed, and then extended to N dimensions by appending it as the last dimension. The resulting N-dimensional kernel has a specific asymmetry, where one side of the last dimension is populated with arbitrary values and the other side is zeroed out. This technique is particularly useful in convolutional neural networks, image processing, or other domains where kernel design impacts computational efficiency and performance.
9. The computer implemented method according to claim 1 , wherein the representation is modified for image transformation, and/or for image recognition, and/or for anomaly detection and/or for image validation.
This invention relates to computer-implemented methods for processing digital images, particularly for enhancing or analyzing image data. The method involves generating a representation of an image, which can be modified for various purposes, including image transformation, image recognition, anomaly detection, and image validation. Image transformation may involve altering the image's appearance or structure, such as resizing, rotating, or applying filters. Image recognition involves identifying objects, patterns, or features within the image. Anomaly detection focuses on identifying irregularities or deviations from expected patterns. Image validation ensures the integrity and correctness of the image data. The representation may be adjusted dynamically based on the specific application, allowing for flexible and adaptive image processing. This method improves efficiency and accuracy in tasks requiring image analysis or manipulation, addressing challenges in automated image interpretation and quality assurance. The approach is applicable in fields such as computer vision, medical imaging, surveillance, and quality control, where reliable and adaptable image processing is essential.
10. The computer implemented method according to claim 1 , wherein an at least partial autonomous vehicle or robot is controlled depending on the representation.
Autonomous vehicles and robots require precise control systems to navigate environments safely and efficiently. A key challenge is accurately representing and interpreting environmental data to make real-time decisions. This invention addresses this problem by controlling an autonomous vehicle or robot based on a generated representation of its surroundings. The representation is derived from sensor data, such as LiDAR, cameras, or radar, which is processed to identify and classify objects, obstacles, and navigable paths. The system uses this representation to determine optimal control actions, such as steering, acceleration, or braking, ensuring safe and efficient operation. The method may also incorporate machine learning models to improve accuracy over time by learning from past interactions and environmental changes. By dynamically adjusting control parameters based on the representation, the system enhances adaptability in varying conditions, reducing the risk of collisions and improving navigation efficiency. This approach is particularly useful in dynamic environments where real-time decision-making is critical.
11. A computer implemented method for digital video enhancement, in which each element of a plurality of elements representing a pixel of a digital image of a digital video includes an indication of a spatial dimension, the spatial dimension indicating a position of the pixel in the digital image, and the element includes an indication of a channel dimension, the channel dimension indicating a channel of the pixel in the digital image and an indication of a time dimension, the time dimension, indicating a position of the digital image in a video timeline of the digital video, the method comprising the following steps: transforming the plurality of elements representing pixels of the digital image to a representation depending on an invertible linear mapping, wherein the invertible linear mapping maps an input of the plurality of elements to the representation; modifying the representation to determine a modified representation depending on the representation; determining a plurality of elements representing pixels of an enhanced digital video depending on the modified representation; and transforming the modified representation depending on an inversion of the invertible linear mapping; wherein the invertible linear mapping includes at least one autoregressive convolution.
This invention relates to digital video enhancement, specifically improving video quality through a multi-dimensional transformation process. The method addresses the challenge of enhancing video frames while preserving spatial, temporal, and channel relationships between pixels. Each pixel in a video frame is represented by an element containing spatial (position in the frame), channel (color or intensity), and time (frame position in the video timeline) dimensions. The method processes these elements by first applying an invertible linear mapping to transform the pixel data into a modified representation. This transformation uses at least one autoregressive convolution, allowing dependencies between pixels to be captured efficiently. The transformed representation is then modified to enhance video quality, such as reducing noise or improving sharpness. The modified representation is inverted back to the original pixel format, producing an enhanced digital video. The use of autoregressive convolutions ensures that the enhancement process maintains coherence across spatial, temporal, and channel dimensions, resulting in visually improved video output. The method is fully reversible, allowing for precise control over the enhancement process.
12. A computer implemented method for digital audio enhancement, in which each element of a plurality of elements representing a part of a digital audio sample includes an indication of a spatial dimension, the indication of the spatial dimension is a constant value, and the element includes an indication of a time dimension, the time dimension indicating a position in an audio timeline of the audio sample, the method comprising the following steps: transforming the plurality of elements representing parts of the audio sample to a representation depending on an invertible linear mapping, wherein the invertible linear mapping maps an input of the plurality of elements to the representation; modifying the representation to determine a modified representation depending on the representation; determining a plurality of elements representing parts of an enhanced digital audio sample depending on the modified representation; and transforming the modified representation depending on an inversion of the invertible linear mapping; wherein the invertible linear mapping includes at least one autoregressive convolution.
This invention relates to digital audio enhancement, specifically improving audio quality by transforming and modifying audio data in a structured way. The method processes digital audio samples where each element includes spatial and time dimensions. The spatial dimension is a constant value, while the time dimension indicates the position in the audio timeline. The method involves transforming the audio elements into a new representation using an invertible linear mapping, which includes at least one autoregressive convolution. This mapping ensures that the transformation can be reversed, preserving the original structure. The transformed representation is then modified to enhance the audio, producing a modified representation. From this, a new set of audio elements is generated, representing an enhanced version of the original audio sample. The modified representation is then transformed back using the inverse of the original mapping, reconstructing the enhanced audio in its original format. The use of autoregressive convolution in the mapping allows for efficient and reversible processing, ensuring high-quality audio enhancement while maintaining the integrity of the original signal. This approach is particularly useful for applications requiring real-time or high-fidelity audio processing, such as music production, speech enhancement, or noise reduction.
13. The computer implemented method according to claim 12 , wherein the constant value is one.
This invention relates to computer-implemented methods for processing data, specifically focusing on optimizing calculations involving constant values. The problem addressed is the computational inefficiency that arises when processing operations with constant values, particularly when those values are frequently reused in calculations. The invention improves efficiency by recognizing and handling constant values in a specialized manner, reducing redundant computations and improving performance. The method involves identifying a constant value within a set of data or operations. Once identified, the constant value is processed in a way that minimizes redundant calculations. For example, if the constant value is one, the method may simplify operations involving this value, such as multiplication or division, to avoid unnecessary computations. The approach ensures that the constant value is used optimally, reducing the overall computational load and improving processing speed. The invention is particularly useful in applications where repetitive calculations are common, such as in numerical simulations, data processing pipelines, or real-time systems where performance is critical. By optimizing the handling of constant values, the method enhances efficiency without altering the accuracy of the results. The technique can be applied across various computing environments, including general-purpose processors, specialized hardware accelerators, or distributed computing systems.
14. The computer implemented method according to claim 12 , wherein the digital audio sample includes audio channels, wherein each element of the plurality of elements includes an indication of a channel dimension, the channel dimension indicating an audio channel in the audio sample, and the plurality of elements including the indication of the channel dimension and representing parts of the audio sample is transformed to the representation depending on the invertible linear mapping, wherein the invertible linear mapping maps an input of the plurality of elements comprising the indication of the channel dimension to the representation, wherein the representation is modified to determine the modified representation depending on the representation, and wherein the plurality of elements comprising the indication of the channel dimension and representing parts of an enhanced digital audio sample is determined depending on the modified representation, wherein the modified representation s transformed depending on the inversion of the invertible linear mapping.
This invention relates to digital audio processing, specifically methods for transforming and enhancing audio signals using invertible linear mappings. The problem addressed is the efficient representation and modification of multi-channel audio data while preserving the ability to reconstruct the original or enhanced audio with high fidelity. The method processes a digital audio sample containing multiple audio channels. Each element in the audio data includes a channel dimension indicator, specifying which audio channel it belongs to. These elements are transformed into a representation using an invertible linear mapping, which preserves the relationship between the input elements and their transformed form. The representation is then modified to achieve desired enhancements, such as noise reduction or dynamic range adjustment. The modified representation is transformed back into the original domain using the inverse of the linear mapping, producing an enhanced digital audio sample that retains the original channel structure. The key innovation is the use of channel-specific transformations that allow for precise modifications while ensuring perfect reconstruction of the enhanced audio. This approach is particularly useful in applications requiring high-quality audio processing, such as real-time audio enhancement or multi-channel audio editing. The method ensures that modifications are applied consistently across all channels, maintaining spatial and temporal coherence in the output.
15. A computer implemented method for encoding digital audio data, in which each element of a plurality of elements representing a part of a digital audio sample includes an indication of a spatial dimension, wherein a first indication and a second indication of the spatial dimension is a constant value, wherein the element includes an indication of a time dimension, the time dimension indicating a position in an audio timeline of the audio sample, the method comprising: transforming the plurality of elements representing parts of the audio sample to a representation depending on an invertible linear mapping, wherein the invertible linear mapping maps an input of the plurality of elements to the representation; and transmitting or storing the representation; wherein the invertible linear mapping includes at least one autoregressive convolution; wherein the digital audio sample includes audio channels, wherein each element of the plurality of elements includes an indication of a channel dimension, the channel dimension indicating an audio channel in the audio sample, and the plurality of elements including the indication of the channel dimension and representing parts of the audio sample is transformed to the representation depending on the invertible linear mapping, wherein the invertible linear mapping maps an input of the plurality of elements including the indication of the channel dimension to the representation, and wherein the representation is transmitted or stored.
This invention relates to digital audio encoding, specifically a method for compressing or transforming audio data while preserving spatial and temporal information. The method processes digital audio samples structured as a plurality of elements, where each element includes spatial, temporal, and channel dimensions. The spatial dimension is represented by constant values, indicating fixed spatial relationships across the sample. The time dimension specifies the position of each element within the audio timeline, while the channel dimension identifies the audio channel (e.g., left, right, or multi-channel configurations) to which the element belongs. The method transforms these elements into a compressed or encoded representation using an invertible linear mapping, which ensures lossless reconstruction of the original audio. This mapping incorporates at least one autoregressive convolution, a technique that leverages dependencies between adjacent elements to improve compression efficiency. The transformed representation is then transmitted or stored. The encoding process maintains the spatial and channel dimensions, ensuring that multi-channel audio (e.g., stereo or surround sound) is accurately represented. The use of autoregressive convolution allows for efficient encoding while preserving the integrity of the original audio data. This approach is particularly useful in applications requiring high-quality audio compression, such as streaming, storage, or real-time audio processing.
16. The computer implemented method as recited in claim 15 , wherein the constant value is 1.
This invention relates to a computer-implemented method for processing data, specifically addressing the challenge of efficiently handling numerical computations in digital systems. The method involves determining a value based on a mathematical operation and then applying a constant value to adjust or modify the result. The constant value is fixed at 1, ensuring consistency in the computation process. This adjustment step is critical for maintaining accuracy and reliability in subsequent data processing tasks. The method is particularly useful in applications where precise numerical operations are required, such as in scientific computing, financial modeling, or data analysis. By incorporating this fixed constant, the method ensures that the output remains stable and predictable, reducing errors that could arise from variable adjustments. The technique is integrated into a broader system that performs iterative or recursive computations, where maintaining a consistent baseline is essential for convergence and accuracy. The use of a constant value of 1 simplifies the implementation while ensuring that the computational results are both accurate and reproducible. This approach is beneficial in environments where computational efficiency and reliability are paramount, such as in real-time data processing or high-performance computing scenarios.
17. A device, comprising: a processor; and storage comprising instructions for a convolutional neural network; wherein the processor is configured for digital image enhancement, in which each element of a plurality of elements representing a pixel of a digital image includes an indication of a spatial dimension, the spatial dimension indicating a position of the pixel in the digital image, and the element includes an indication of a channel dimension, the channel dimension indicating a channel of the pixel in the digital image, the processor configured to: transform the plurality of elements representing pixels of the digital image to a representation depending on an invertible linear mapping, the invertible linear mapping mapping an input of the plurality of elements to the representation; modify the representation to determine a modified representation depending on the representation; determine a plurality of elements representing pixels of an enhanced digital image depending on the modified representation; and transform the modified representation depending on an inversion of the invertible linear mapping, wherein the invertible linear mapping includes at least one autoregressive convolution.
This invention relates to digital image enhancement using a convolutional neural network (CNN) with an invertible linear mapping. The problem addressed is improving image quality while preserving structural integrity and computational efficiency. The device includes a processor and storage containing CNN instructions. The processor processes digital images where each pixel is represented by elements containing spatial (position) and channel (color or intensity) dimensions. The processor transforms the pixel elements into a representation using an invertible linear mapping, which includes at least one autoregressive convolution. This mapping ensures that the transformation can be reversed, preserving the original image structure. The representation is then modified to enhance the image, and the modified representation is transformed back into pixel elements using the inverse of the original mapping. The result is an enhanced digital image. The use of autoregressive convolutions allows the network to model complex dependencies between pixels while maintaining invertibility, enabling high-quality enhancements without losing critical image information. This approach is particularly useful for applications requiring both high fidelity and computational efficiency, such as medical imaging or real-time video processing.
18. The device according to claim 17 , further comprising an output adapted to output a result of an image transformation, an image recognition, an anomaly detection and/or an image validation.
This invention relates to a device for processing and analyzing images, particularly in applications requiring image transformation, recognition, anomaly detection, or validation. The device is designed to enhance the accuracy and efficiency of image-based systems by performing advanced computational tasks on input images. The core functionality includes transforming images to improve their quality or extract specific features, recognizing objects or patterns within the images, detecting anomalies or deviations from expected norms, and validating the integrity or authenticity of the images. The device integrates these capabilities to provide a comprehensive solution for image analysis, ensuring reliable and actionable results. By outputting the results of these processes, the device enables real-time decision-making in applications such as surveillance, medical imaging, industrial inspection, and autonomous systems. The inclusion of multiple analysis functions within a single device streamlines workflows and reduces the need for separate systems, improving overall system performance and reducing latency. The device is particularly useful in environments where rapid and accurate image interpretation is critical, such as in automated quality control, security monitoring, or diagnostic imaging. The output of the device can be used to trigger further actions, generate alerts, or provide data for downstream processing, making it a versatile tool for various image-based applications.
19. The device according to claim 17 , further comprising an actuator adapted to control an at least partial autonomous vehicle or robot depending on the representation, and/or depending on a result of processing the representation, and/or depending on image data determined by the inversion of the invertible linear mapping.
This invention relates to systems for processing and utilizing image data, particularly in autonomous vehicles or robots. The core problem addressed is the efficient and accurate interpretation of visual data to enable autonomous decision-making. The device includes a processor configured to generate a representation of image data by applying an invertible linear mapping, such as a Fourier transform or wavelet transform, to the image data. This representation is then processed to extract relevant features or information. The device further includes an actuator that controls the autonomous vehicle or robot based on this representation, the processed results, or the original image data reconstructed from the invertible linear mapping. The actuator ensures that the vehicle or robot responds appropriately to the environment by adjusting its movements or actions according to the analyzed data. This approach enhances the system's ability to interpret complex visual information and make real-time decisions, improving autonomy and safety in dynamic environments. The use of invertible linear mappings allows for efficient data compression and reconstruction, optimizing computational resources while maintaining accuracy.
20. A non-transitory computer-readable medium on which is stored instructions for digital image enhancement, in which each element of a plurality of elements representing a pixel of a digital image includes an indication of a spatial dimension, the spatial dimension indicating a position of the pixel in the digital image, and the element includes an indication of a channel dimension, the channel dimension indicating a channel of the pixel in the digital image, the instructions, when executed by a computer, causing the computer to perform the following steps: transforming the plurality of elements representing pixels of the digital image to a representation depending on an invertible linear mapping, the invertible linear mapping mapping an input of the plurality of elements to the representation; modifying the representation to determine a modified representation depending on the representation; determining a plurality of elements representing pixels of an enhanced digital image depending on the modified representation; and transforming the modified representation depending on an inversion of the invertible linear mapping, wherein the invertible linear mapping includes at least one autoregressive convolution.
This invention relates to digital image enhancement using a transformative approach. The problem addressed is improving image quality while preserving structural integrity. The method involves processing a digital image represented by pixel elements, each with spatial (positional) and channel (color/intensity) dimensions. The core process includes transforming the pixel data into a new representation using an invertible linear mapping, which includes at least one autoregressive convolution. This transformation allows for efficient manipulation of the image data. The transformed representation is then modified to enhance the image, and the modified representation is converted back to pixel data using the inverse of the original mapping. The autoregressive convolution ensures that the transformation is both invertible and preserves spatial relationships within the image. This approach enables high-quality enhancements while maintaining computational efficiency and reversibility. The method is particularly useful for applications requiring real-time image processing, such as medical imaging, surveillance, and augmented reality.
Cooperative Patent Classification codes for this invention. Click any code to explore related patents in that topic.
December 3, 2019
March 15, 2022
Browse 5M+ US patents with plain-English claim translations and AI-generated analysis.